Official project status · evidence-first
Ounce-Core: current status and evidence
This page is the maintained entry point for readers who find old launch-day articles or mirrored snapshots. The project changed direction on 8 November 2026: see the vision. Check linked GitHub records for the latest run rather than treating an old figure as current.
What has progressed since the October 2026 launch?
- Differential tests against real hardware: LZ4, xxHash, SQLite 3.50.4, Lua 5.4.7 and a 5,964-line floating-point instruction test match an Apple Silicon reference byte for byte on Tier-3, Tier-4 and native. Method and results: docs/verify. Native runs on the host CPU, so Tier-3 and Tier-4 are the real evidence. More programs are in review.
- Failure-to-reproduction toolkit (
ounce-diagnosev0.2.0): shipped as a standalone Python archive (diagnosis-v0.2.0 release, CLI guide) packaging the retained SDL2 frame-divergence walkthrough and the external libnx #685 audio-channel padding case. - Bugs found and fixed: a prefetch instruction decoded as a load, wrong-sign fused floating-point instructions (wrong
sin,cos,pow), and missing vector instructions (widening, narrowing, saturating, byte-reverse, table lookup, polynomial multiply, register shifts), each with unit tests. - Robustness: 300 corrupted programs through the loader produced no panics, no host signals and no hangs on the interpreter engine. Native execution was not fuzzed. See stress results (once merged).
- Legacy corpus: 50 pinned Switch-era homebrew programs still pass their first-frame checks in CI. They are a regression set, not the product direction.
- Switch 2 work: parked per PLAN.md (existing checks retained; no coverage expansion without a demonstrated user need). It was never evidence of commercial software support.
9 November 2026: newer bounded milestones
- Static Linux AArch64 ELF: A restricted
run-elfpath is merged (#255), with pinned LZ4, xxHash, cJSON and zlib static musl ELFs verified across Linux, macOS and Windows runners (#277, #292). On 10 November, 18 pinned programs in total were run as static ELFs against a host build, qemu-user and Tier-3/Tier-4 on Linux (list and limits); wasm3 does not run yet (#366). That run found and fixedLDPSWsign extension and two missing scalar FP forms. This does not imply general Linux application compatibility. - Floating-point control register: FPCR rounding mode, flush-to-zero and default-NaN are now honoured by common scalar and vector FP arithmetic (#363); on the bounded fp-edge corpus, interpreter, Tier-3 and Tier-4 matched Linux Arm and Apple M1 hardware at all nine FPCR values with 0 mismatches (CI run). The interpreter and Tier-3 also raise the FPSR exception flags (#373); Tier-4 compiled FP cannot, and FPCR trap enables and half-precision arithmetic are still not modelled.
- Interactive lab: the web verification workbench was merged (#261). It is a bounded inspection and evidence tool, with an in-browser WebAssembly AArch64 CPU stepper (/playground/).
- Differential instruction fuzzer: landed in CI (#258), with commutative vector floating-point NaN operand order pinned in #272. FPSR high-bit retention on Neoverse remains documented in issue #257.
- Outside replication: no accepted independent run is recorded in the limitations register; interested testers can use the rerun guide and issue #248.
Read primary evidence, not a frozen milestone
- Legacy Switch corpus workflow — run status and downloadable evidence when available
- Living development plan — milestone links and explicit limitations
- Official releases — versioned announcements and assets
- Pull requests — unmerged work is not a shipped capability
- Independent Tester Challenge — outside reports are counted separately, not folded into internal CI metrics
Corrections for old search results
The launch-day state is part of the history, but it is not a description of the current code. An early article may be accurate about what was present in October while omitting subsequent releases, testing and architecture changes. This page does not demand removal of historical criticism or claim that newer work retrospectively validates earlier assertions.
For researchers and journalists: cite the relevant commit SHA, release, workflow run ID, registered oracle and execution engine. Distinguish an internally reproduced test from an independent outside rerun. Please link directly to this page and the primary artifact when reporting new results.
Editorial review: 9 November 2026 (review of merged PR scope and recorded limitations; consult CI for run-specific claims).